How can lawyers use AI without breaking professional rules?

Generative AI can save a lawyer real time: turning a long transcript into a summary, suggesting the structure of a memo, or rewriting a dense paragraph for a client who isn't a lawyer. The professional duties that apply to every other part of practice apply here too. The guidance published so far by bars and courts says, in effect, that the tool is allowed but the lawyer stays fully responsible for what goes in and what comes out.

This article explains that guidance for a general audience. It is not legal or ethics advice. Rules differ by jurisdiction, so every lawyer should check what their own regulator, whether a law society, bar or court, has published.

The warning everyone remembers

In a US federal case reported in May 2023, lawyers suing the airline Avianca had filed a brief that cited six court decisions that did not exist. ChatGPT had made them up. When one lawyer asked the chatbot whether a case was real, it insisted it was and said it could be found in standard legal databases. The judge described the cases as bogus decisions with bogus quotes and scheduled a hearing on sanctions.

The lesson is not that AI is forbidden. It is that a chatbot produces text that sounds right, including invented case names, citations and quotations, and asking the same chatbot to confirm its own answer is not verification.

What the guidance says

In November 2023, the State Bar of California's committee on professional responsibility issued practical guidance, framed as guiding principles rather than best practices. Its main points:

• Confidentiality. Don't put confidential client information into a tool that lacks adequate confidentiality and security protections. Read the terms of use to learn how inputs are used, anonymize client details, and consult IT or security experts before entrusting a system with client data.
• Competence. Understand how the tool works and its limits. AI output can be a starting point, but it must be critically reviewed and corrected, and professional judgment can't be delegated to it.
• Candor to the court. Check all output, including citations, before filing, and check local rules on disclosing AI use.
• Billing. Charging for time actually spent writing prompts and reviewing output is fine; billing hours the AI saved is not.
• Supervision and communication. Firms should set policies and train staff, and lawyers should consider telling clients how they plan to use AI.

In Canada, the Federal Court issued a notice on December 20, 2023. If a document filed in litigation contains content generated by AI, its first paragraph must say so. The Court urges caution with AI-generated legal references, recommends relying on established sources such as official court websites or CanLII, and calls for a "human in the loop" to verify AI-created content. The notice applies to lawyers and self-represented parties alike.

A worked research-and-draft workflow

Here is one cautious way to use a general chatbot on a research memo, following the duties above.

1. Check the rules first. Your regulator's guidance, the court's practice directions and the client's own instructions on AI.
2. Strip identifying details. Replace names, dates and unique facts with neutral placeholders.
3. Ask for structure, not authority. For example:

I am preparing a research memo on whether a residential landlord can keep a security deposit for normal wear and tear in [JURISDICTION]. Do not cite any cases or statutes. Instead:
1. List the legal questions I need to answer, in a logical order.
2. For each, list the kinds of sources I should look for.
3. List the facts I would need from the client.

1. Do the research in trusted sources. Find the actual statutes and decisions in official or established legal databases, and read them.
2. Draft with the tool if useful, pasting in only verified authorities and anonymized facts.
3. Verify everything again. Every citation, quotation and statement of law, against the original.
4. Record and disclose as required. Note how AI was used, add any declaration the court requires, and bill only for time actually spent.

What must never go into a public chatbot?

Anything confidential about a client should stay out of a public consumer chatbot unless you have confirmed, from its terms and with expert advice, that it offers adequate confidentiality and security, which a free general-purpose service usually can't be assumed to do. That includes client names and anything that could identify them, the facts of their matter, privileged communications, and documents shared with you in confidence.

Some information is off-limits for additional reasons: material covered by a court order or protective order, documents obtained in discovery, and personal information about other parties, witnesses or employees. Business details such as unannounced deals, settlement figures and negotiation positions also belong outside.

When in doubt, a practical test helps: if you wouldn't read it aloud in a crowded café, don't paste it into a public chatbot. Anonymize thoroughly, ask the general question instead of the client's specific one, and use tools your firm has vetted for anything that needs real client data.